
THEWHITEBOX
TLDR;
Welcome back! This week is packed with model releases, including two from OpenAI and Anthropic that dropped today. This past week also saw the launch of the first decision model, Jev, plus new Neuralink trials, market data, and products, so there’s a lot to talk about.
Must-read, in my opinion.
Enjoy!

RESEARCH
Meet Jev, the first decision model

An alternative to LLMs has arrived… although not quite, as you’re about to see. In reality, it provides a solution to a problem, automation, that LLMs have been consistently used for and never should have.
I wrote a detailed piece on Medium if you want to read it for free.
This concept is being defined as we speak, but I bet they will be called decision models, because that’s literally all they do: you give them a problem that requires a decision, and they make it.
Classifying emails, routing support tickets (as in the thumbnail), filtering tables… that’s the kind of work decision models will do.
But don’t LLMs already solve this? Well, not quite.
LLMs are too slow, expensive, non-deterministic, and non-calibrated (they consistently give overconfident responses), making them quite suboptimal solutions for automation.
Jev doesn’t generate text, images, or control humanoid minds. But it makes decisions correctly, fast, and consistently.
The price is also incredibly competitive: $0.042 per million input tokens, with output tokens literally free because there are so few that it’s too cheap to meter. This means this solution is hundreds of times cheaper than frontier models.
And did I say it’s fast? Because it’s really fast. It can output tens of decisions per second in the same period an LLM makes one. Most interestingly, decision models can be used in truly unexpected ways.
Just to show a few:

Jev playing Doom. Source: TypeSafe
An emoji searcher that shows you which emojis match your search. Paired with embeddings, this could build real recommendation systems that are fast enough and smarter than the average reranker. In fact, I believe we’ll have decision models in any product that requires search or filtering in no time.

Source
Finding Zürich → London flights on Google Flights in 7.1 seconds for $0.0039. The video is at 1× speed. You can watch the agent click, type, and land on results at incredible speed.
Classifying 500 emails in a few seconds for 3.5 cents — labels appear on screen as it runs. A longer run does 1,700 emails for 18 cents.
Playing Super Mario Bros. from an emulator state. The post is the gameplay, not the source tree.
Labeling 1,018 AI papers for $0.08 at 256 ms each, then browsing the map at 1kpapers.com.
Tearing down 724 live ads from 37 brands in 40 seconds for 9 cents — hook, format, offer, CTA, awareness stage, landing-page mismatch.
Putting
jev()inside Postgres:WHERE jev(people, 'could work from home'). You can use Jev to create filters in your databases. No index, no embeddings. The clip shows 129 rows judged in ~1 s for $0.0009. This is my favorite example because it has immediate application in thousands of software applications.
TheWhiteBox’s takeaway:
Some of the use cases are incredibly cool. I’m incredibly excited about decision models because I’ve been notoriously critical of the idea that LLMs can be used reliably for automation.
They cannot. They are not reliable at all, and too slow and expensive despite that lack of robustness. They were begging to be disrupted, and they got disrupted.
To be clear, this is not a mind-bending new approach, technologically speaking; it’s a machine learning classifier, a decades-old idea. What makes this one special is that it is general, so that you don’t have to fine-tune your classifier for every task. Instead, just drop Jev in there and it will work.
I don’t know whether Jev and TypeSafe will be great investments for their founders and backers. I don’t know if Jev wins the category (probably not).
But this is clearly a category-defining model that I believe will be extremely successful.
BCIs
Giving a person’s voice back
Neuralink released a demonstration on September 18 showing Terry, a participant in its VOICE clinical trial, using a brain implant to produce speech in a recreation of his own voice.
Terry was diagnosed with bulbar-onset amyotrophic lateral sclerosis (ALS) in 2024. He can still move his mouth and body, but his speech has deteriorated to the point that his loved ones struggle to understand him.
To train the system, Terry first mimed speech as best he could, helping the algorithm learn to associate his brain signals with the words he intended to say. Neuralink said he then progressed to thinking the words and hearing them spoken aloud in his recreated voice. The demonstration forms part of the company’s work on a brain-to-voice interface for people with severe speech impairment.
TheWhiteBox’s takeaway:
The devices remain investigational and have not received regulatory approval, but boy is this exciting, and all about what AI should be. A solution to make human lives better, not worse, and applications free of fearmongering, regulatory capture, and straight negativity.
AI has a bad rep because of LLMs, but it’s way more than just that. Ironically, BCIs are surprisingly similar in what they do: they take an input token sequence, in this case brain signals encoded into tokens, and they capture the patterns in the data that allow us to decode what the human is thinking.
In other words, the model is fed an enormous amount of “if the human thinks of saying ‘x’, the brain patterns look like ‘y’”, and it learns to capture those patterns so that whenever the AI sees the brain patterns of the user, it can say, “ok, the user wants to say he’s looking forward to lunch.”
The caveat is that brain patterns don’t seem very generalizable right now, meaning how you and I think about having lunch produces considerably different brain patterns, so a Neuralink in your brain won’t work in mine without some fine-tuning.
Poor generalization is clearly the main remaining hurdle preventing Neuralink from potentially curing things like voice, sight, or human-computer interaction. I don’t think it’s a technological issue at this point, but one of scale and economic feasibility, since I assume fine-tuning the model to every single human isn't particularly scalable (besides the fact that Neuralink is an invasive chip that requires surgical implantation).
Still, the fact that we’re already giving a voice to people who lost it, or to completely paraplegic humans interacting with computers as a form of communication with the world, is not only beautiful, but a breath of fresh air at a time when AI is almost synonymous with potential negative outcomes.
CYBERSECURITY
There’s an “Irregular” pattern in AI cyber incidents
A few days ago, Google confirmed that its Gemini AI model accessed three companies’ systems without authorization during cybersecurity tests in May. External evaluator Irregular conducted the tests, according to The Wall Street Journal, which first reported the incidents.
Google said Gemini used publicly available information to guess or retrieve login credentials for websites it believed belonged to the test. The company said the model stopped in all three cases. Google’s vice-president of security engineering, Heather Adkins, said it had notified the affected businesses and worked with its testing partner to revise its procedures, Reuters reported.
Interestingly, unlike what OpenAI or Anthropic love to infer from their models, Google said it did not consider the incidents evidence of model misalignment, meaning behavior departing from its intended objectives.
TheWhiteBox’s takeaway:
What’s interesting is that every “bad event” from these labs traces back to the same company: Irregular.
Irregular, formerly Pattern Labs, is an AI cybersecurity company that tests how capable advanced models are at hacking and develops ways to secure those models. It has worked with OpenAI, Anthropic, Google DeepMind, and the UK’s AI Security Institute.
A central part of its work involves giving AI models cybersecurity challenges inside simulated environments: can they discover vulnerabilities, exploit them, bypass defenses, and complete an attack autonomously? This helps AI labs measure their models' capabilities and risks. OpenAI has included Irregular’s evaluations in its model safety reports.
It also works to protect AI systems, including preventing theft of model weights and securing inference so sensitive information stays confidential. Think of it as a specialist security testing lab and defense builder for frontier AI: it investigates both what attacks AI can carry out and how AI systems can be attacked.
And well, it seems Irregular has some irregular security practices.
RESEARCH
There’s an irregular pattern in AI cyber incidents

Anthropic says Claude now leads 26% of its AI research and development work, up from less than 1% in February 2026. Its new measurements of internal AI development show that, by August, more than 90% of measured work involved substantial AI collaboration or greater automation. No measured category was fully autonomous though.
The index uses an automation scale developed by Epoch AI. “AI leads” means a system can complete most of a task from a high-level instruction while a human supervises. Anthropic used Claude to classify internal work and assess automation, acknowledging that using its own models as evaluators could produce shared errors.
About 30,000 research and engineering agents were running simultaneously on Anthropic’s most-used internal platform in August. The company reported that every action passed through an automated monitor before execution; approximately 0.002% of more than a billion decisions were blocked.
TheWhiteBox’s takeaway:
Using AIs to push AI research forward is not an Anthropic-only reality. A few days ago, Z.ai, the company behind the Chinese GLM models, gave an actual example of how its AIs are helping with AI work.
Z.ai said the GLM-5.3 model helped engineers build the software infrastructure now serving GLM-5.3-Flash on more than 100,000 Chinese-made AI accelerators (chips). According to the company’s technical account, the work progressed from initial adaptation to production readiness in less than two weeks, tripling throughput against the starting baseline.
An agent powered by GLM-5.3 helped diagnose software faults and optimize the code running on the chips. The team supplied tests and performance measurements so the agent could check its proposed changes and revise them. Engineers and the agent jointly optimized the system, with correctness, stability, and overall performance determining whether they accepted changes.
The deployment faced limited chip memory and bandwidth, incomplete software support, and the need to handle multimodal requests and a million-token context window. Z.ai said memory optimizations and changes to how processing stages were distributed brought hardware efficiency and per-token costs to levels comparable with mainstream Nvidia GPUs. Those comparisons are company-reported results with little proof, so we have to take their word for it.
The problem is that this example from the Chinese, and those Anthropic allude to, “suffer” from the same virtues that make AI research automation very possible but not very applicable outside the domain, because:
It’s heavily experimentation-based, so an AI using a lot of compute to try every single possibility and seeing what sticks is a genuine research strategy
The domains are verifiable; we can instantly tell whether the new solution is better than the previous one, at least from a benchmarking perspective. This allows the AIs to ‘hillclimb’ their way into better solutions, with ‘time’ being the only meaningful variable separating you from your next improvement
It has a lot of dollars behind it. To deploy an enormous amount of compute into a problem, you need enormous amounts of cash. This only makes sense in areas where the potential economic return justifies it.
Unbound by reality. Progress isn’t bottlenecked by physical constraints, as we see in areas like drug discovery (e.g., human trials)
Most real-world applications don’t meet any of these three conditions, let alone all four, so we must be very careful about taking these results and extrapolating them to ‘ah, yeah, next week we’ll cure cancer’. Well, no.
These results do show greater value of AI tools, but this industry loves to extrapolate and jump to conclusions faster than we should.

START-UPS
Harvey & an Important Lesson on AI margins
Harvey is turning to open-weight AI to reduce its dependence on OpenAI and Anthropic after rising usage drove its gross margin below zero. Bloomberg reported that the legal software company’s margin fell from about 50% at the start of 2026 to −50% in June following a March update to its AI agents. Harvey said its token consumption had increased twentyfold this year, but declined to comment on the financial figures.
The company introduced Tenet in an August 20 research preview. Built on Moonshot AI’s Kimi K3 and post-trained with Fireworks, the model is designed for legal assignments requiring extended, multistep work. Harvey said its initial results showed improvements in performance and cost efficiency.
Bloomberg reported that the model and other changes to Harvey’s AI usage had helped restore positive gross margins, citing people familiar with the effort.
TheWhiteBox’s takeaway:
The shift does not amount to abandoning proprietary models. Harvey continues to offer OpenAI and Anthropic systems, with its September 18 release notes adding access to GPT-6 Astra and Fable 5.1.
However, at this point, the writing is on the wall: you don’t need frontier tokens for most features.
That is why running your business on frontier lab models is so dangerous: they are selling API models at great margins, which severely limits your ability to make money, so you end up eating the cost entirely. Thus, it’s not surprising that Harvey turned into fine-tuned open models; it’s the only way the numbers make sense!
Does this mean frontier AIs no longer make sense for these application-layer companies to use? Hardly, they have a role to play on those long-horizon tasks that require higher “intelligence” and which Harvey can probably charge for a premium.
But using them for your standard product features is committing margin seppuku. But this news has another big lesson: software margins have changed.
Put another way, this is a beautiful exemplar of the wild differences between non-AI software and AI-powered software when it comes to margins.
In a non-AI world, what happened to Harvey would not have happened in a million years. Non-AI software is not only much more predictable in terms of hardware intensity (i.e., how many chips you need per workload is considerably predictable), but that guarantees that marginal costs are negligible; every new user you add to the platform results in ~0 additional serving costs because the hardware intensity of that new user is minimal (i.e., your cluster can handle it and most part of the added costs are minimal).
With AI, an employee/customer with excessive ambition can effectively hijack an entire server with a single sequence, skyrocketing cost/token and forcing you to charge an amount you can’t usually charge, automatically sending you into red territory because you have to eat the cost.
This is the main reason I still believe AI subscriptions don’t make sense. One thing is to have low-usage users subsidizing the power users. Another is the power user consuming the equivalent of 100s of low-usage users.
With AI, this golden rule that, for decades, guaranteed 90%+ gross margins for software companies is broken, because individual users have an enormous “control” over how many chips you’re going to need. This should inform every product decision when you add AI features.
PUBLIC MARKETS
Anthropic Delays the IPO (Again)
Anthropic is targeting a November initial public offering, a month later than many investors expected, The Wall Street Journal reported, citing people familiar with the matter.
Some advisers said the later date would allow the Claude developer to present third-quarter financial results demonstrating its competitive position following OpenAI’s September launch of Astra.
The timetable was chosen before the recent debate over slowing AI development, according to PYMNTS’ account of the Journal’s reporting. The offering could raise as much as $100 billion and value Anthropic at roughly $2 trillion, although the timing remains subject to change. Anthropic did not immediately respond to PYMNTS’ request for comment.
Anthropic announced a confidential IPO filing on June 1, saying the offering would depend on market conditions and other factors. Reuters had reported on September 4 that marketing would begin no earlier than mid-October, with a listing potentially occurring shortly before November’s US midterm elections.
TheWhiteBox’s takeaway:
The question is: will they actually IPO?
My doubts about this entire thing keep growing, because why would they delay the IPO if they are allegedly racing ahead of everybody else?
When you’re so concerned about timing the IPO perfectly, you know something’s off. I don’t know what it is (I can probably guess), but I know something’s off. Waiting a few more weeks to announce that their third quarter was also “adjusted profitable”, a cute way of saying “if you subtract this, this, and that, we’re profitable!” doesn’t sound like a strong reason to delay unless you hold public investors in deep contempt because falling for that trick as an investor is pretty bad.
The truth is, I don’t think a company losing so much money is ready to go public, especially when it claims it wants to slow down and that we should regulate the industry, creating enormous uncertainty. Nonetheless, OpenAI has argued this is the primary reason they are delaying their IPO to 2027.
In my view, this all boils down to one thing: these labs can’t be profitable without a strong regulatory environment protecting them, because the sheer speed of progress forces them into a never-ending spree of new model releases that require new billion-dollar training runs, and while inference may be profitable, it is only as long as you don’t have to train the model, too.
Neocloud/Hyperscaler businesses are great because they don’t have to deal with huge training costs, which is why CoreWeave has positive operating cash flows they can brag about (they still have negative free cash flow because they invest more than they make), and Anthropic has to run marketing campaigns bragging about “adjusted” profits.
We’re treating this as if Anthropic has figured out the business model for AI labs, and it hasn't. Worse, it may only become true if they get the regulatory environment they need, which I don’t think (and trust) they will.
DOOMSAYING
Were Labs Exaggerating?
A September 17 Wall Street Journal opinion column by Brian Gross challenged descriptions of OpenAI’s Hugging Face breach as AI “going rogue,” arguing that testing decisions and weakened safeguards drove the incident.
The column drew on Eryk Salvaggio’s analysis in the Bulletin of the Atomic Scientists, which emphasized human responsibility for the testing environment.
As you probably know already (I explained the incident in detail here), hundreds of OpenAI agents working on coding tasks found a way to communicate without permission and coordinated an attack on Hugging Face to obtain benchmark answers. They shared discoveries and tried to hide some of their actions, with some continuing despite recognizing that the attack fell outside their assigned tasks.
The key point, or crux, of Brian Gross's argument is that agents ran with fewer safeguards than OpenAI’s public products, and that OpenAI’s evaluation infrastructure had severe bugs.
In other words, what he is saying, and I agree, is that while the AI showed some incredible capabilities, the blame is not on the AI; it’s on the humans at OpenAI and, importantly, the AI did exactly what was expected from it.
TheWhiteBox’s takeaway:
The message is clear: regulation has not to be enforced in a way that creates a moat for these labs. Instead, the approach should be much more ‘mundane’: exhaustive liability risks to the labs if their models misbehave.
I particularly like Judge Glock’s approach: tort liability to ‘non-customers’. That is, the liability must be enforced not on the lab customer using lab models to serve others, but on the labs themselves even if the affected aren’t direct customers.
If you enforce this, I can promise you all the “mistakes” occurring from frontier labs’ end. And right quick.

MODELS
OpenAI Releases GPT-6 Sol and GPT-6 Luna

OpenAI has released GPT-6 Sol and GPT-6 Luna, bringing training advances from its flagship Astra model to cheaper models for coding, computer use, and professional work.
Sol costs $2 per million input tokens and $10 per million output tokens.
Luna costs $0.10 and $0.50 per million input and output tokens, respectively.
These prices are at least 50% below their predecessors’ promotional rates. Astra remains the company’s highest-capability model. The benchmarks are similar to those obtained by GPT-5.6 Sol/Luna, but at much lower prices, showing this is mostly a “Pareto release”: improved intelligence per cost.
And based on the Artificial Analysis cost-per-task benchmark, they really up the ante on intelligence per cost, showing how OpenAI basically dominates the curve.

Both models are rolling out in ChatGPT Work, Codex, and the API. Free and Go users can access Luna through the desktop app. Neither model is available in Chat at launch.
TheWhiteBox’s takeaway:
Much like Opus 5.5 below, this clearly signals that OpenAI and Anthropic are feeling downward price pressure.
As for how they manage to offer such big improvements in intelligence per cost, it is because these new models are likely smaller, but as looped transformers, they can perform arbitrarily long computation per prediction without necessarily increasing that computation by growing their size.
In other words, if you have a model with fixed computation per prediction, the only way to increase that computation is by increasing model size or chain-of-thought length.
But with looped transformers, a model with just a single loop has the same computational graph as a sequential one with double the layers that isn’t looping. As labs are now under pressure to drop prices and improve margins ahead of an IPO, looping comes incredibly handy even if that means sacrificing monitorability, as we talked about recently here.
MODELS
Anthropic releases Opus 5.5

Anthropic has released Claude Opus 5.5 today, saying it matches Fable 5.1 on most work while reducing typical workload costs by 40% versus Opus 5 at default settings. The model is available across Anthropic’s platform, AWS, Google Cloud, and Microsoft Azure.
API prices are $4 per million input tokens and $20 per million output tokens, both down 20% versus Opus 5. Cached input costs fall 60%, to $0.20 per million tokens. Anthropic attributes the larger workload savings to cheaper tokens and fewer tokens needed per task, and reports output generation is more than 30% faster.
The release adds safeguards that redirect most cybersecurity tasks to Opus 4.8. Sonnet 5.5 and Haiku 5.5 are scheduled for the coming weeks.
TheWhiteBox’s takeaway:
As seen above, the model is incredibly strong on paper, basically frontier almost everywhere. The question is whether it really is, given that Opus 5 was truly unremarkable despite its incredible benchmark results.
They keep insisting on this rubbish idea of downgrading requests to worse models, something that shouldn’t even be allowed, in my view, but it’s incredibly misleading and extremely hard to predict based on the probabilistic nature of these models.
The obvious takeaway, though, is the price reduction. If Anthropic is reducing prices, you know competition is getting pretty heated. We’ve been calling Anthropic’s pricing “unsustainable” in this newsletter for months now, and we were correct.
If there’s a lesson to be learned here, it is that we must ensure regulation doesn’t prevent these price reductions; if open Chinese models, coupled with Meta/SpaceX, are forcing you to drop prices, so be it; learn to protect pricing beyond models, don’t force the world to use your models just because.
And if you can’t, that’s a you problem, my dear frontier labs.
MODELS
SpaceXAI Launches Grok 4.7

Scores on Cursor Bench (coding benchmark)
SpaceXAI released Grok 4.7 on September 21, targeting coding and professional work with a larger base model and a longer reinforcement-learning run than Grok 4.6. The company says training emphasized harder tasks requiring hours to complete, improving the model’s ability to check its work and manage long contexts. It is available through Cursor, Grok Build, and the API.
In independent testing by Artificial Analysis, Grok 4.7 scored 46 on the Intelligence Index, two points above Grok 4.6, with gains concentrated in professional tasks. Paired with Grok Build, it scored 56 on the Coding Agent Index, up nine points and ranking fourth among models evaluated with their native coding agents, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.
Starting API prices remain $2 per million input tokens and $6 per million output tokens. However, Artificial Analysis measured approximately 81,000 output tokens per task at the new model’s xhigh reasoning setting, compared with 36,000 for Grok 4.6 at high, which has considerably increased cost per task. Its context window remains 500,000 tokens. SpaceXAI also offers a faster variant at twice the price, claiming twice the output speed.
The release includes a redesigned safeguard system that SpaceXAI says improves resistance to jailbreaks while preserving legitimate cybersecurity assistance. Selected cybersecurity partners are receiving invitation-only access to additional red-team capabilities for defensive research.
TheWhiteBox’s takeaway:
SpaceX continues to prove it belongs among the other frontier labs. The acquisition of AnySphere, the company behind Cursor, has aged beautifully. Great acquisition in my opinion. Without them, SpaceX would have been long lost.
However, I think they are making a mistake by not dropping prices to be more competitive like Meta. I know SpaceX doesn’t have Meta’s balance sheet or cash flows to leverage, but I think they should be more aggressive with pricing nonetheless.
FINANCE
Anthropic prepares to launch Claude Money
Anthropic is developing a personal-finance feature called Claude Money that would let users connect bank accounts to its AI assistant, according to TestingCatalog’s September 14 report. The discovery comes from an unreleased interface in Claude’s iOS app, rather than a public product announcement.
The interface includes a separate Money tab and invites users to link accounts and ask Claude about spending and financial plans. The supported banks, financial-data provider, and available actions remain unknown. Subsequent reporting by TechRepublic said Anthropic had not confirmed pricing or a launch date. The disclosed interface does not establish whether Claude could initiate transactions.
TheWhiteBox’s takeaway:
It seems both OpenAI and Anthropic agree that there’s still an opportunity to capture the consumer market, but that can’t be a simple search assistant. They need to offer services that go beyond just answering questions.
The problem with these tools is that they are seldom useful, because they make too many mistakes in situations where they shouldn’t. I believe these solutions should be written in type-safe environments that, ideally, include formal verification to ensure these mistakes aren’t made, with examples like Spark.
I do assume these tools are built on Rust, but I do also believe they are usually rushed. The reasons why they aren’t done this way already are beyond me and possibly because of Silicon Valley’s addiction to moving fast and breaking things.
MODELS
Another massive compression

PrismML released Ternary Bonsai 2 27B, a compressed version of Qwen3.8 27B whose language-model weights occupy about 5.9 GB, compared with roughly 54 GB at 16-bit precision. The company says it retains 98.2% of the original model’s aggregate benchmark performance.
In other words, the model scores almost as well as the original larger one on benchmarks despite being roughly ten times smaller.
The model compresses most language-model weights into three values, −1, 0, and +1, with shared scaling factors. It supports text and image inputs and a context window of approximately 262,000 tokens. According to its model documentation, image processing requires a separate component of about 0.63 GB, beyond the advertised language-model size.
As seen above, despite the aggressive compression, PrismML reported an aggregate score of 83.9 against Qwen3.8 27B’s 85.4 across tests covering reasoning, mathematics, coding, instruction following, vision, and tool use.
TheWhiteBox’s takeaway:
Benchmarks only tell so much. Compression issues might arise in unexpected ways that benchmarks alone may not reveal. Still, sustained benchmark performance is a great sign, and it’s clear we have enormous compression opportunities in AI.
Importantly, our ability to pack greater “intelligence” into smaller “packages” continues to improve at incredible speeds.
But why does this model weigh so much less despite having the same number of parameters? Simple: we use ternary weights.
To translate what that means, let’s look at a simple example. Say a parameter has the value 0.4095. That value must be stored in memory, but memory stores only ones and zeros.
But how many bits do we need to store a number? Take 0.4095. We can write it exactly as 4095 × 10⁻⁴, then store just 4095 and −4, assuming our format already specifies base 10.
Each bit has two possible values: 0 or 1. Adding another bit doubles the combinations available. Eight bits give us 256 combinations, enough to represent integers from 0 to 255. Twelve bits give us 4,096 combinations, covering 0 to 4095.
So 4095 needs 12 bits. The exponent’s magnitude, 4, needs three bits, plus one to indicate its negative sign. That makes 16 bits in this particular format, or 2 bytes.
A precision format defines which values we can represent and how many bits we use to store them. More bits generally let us represent values more precisely, enhancing the model’s ability to learn, but also require more memory.
For example, a model with 10 billion parameters needs 20 GB for its weights in BF16, which uses 16 bits per parameter. Ternary weights have only three possible values. Packing them together can theoretically bring storage down to about 1.58 bits per parameter, or roughly 2 GB for the same parameter count.
The model still has 10 billion parameters, but storing and moving its weights requires about ten times less memory and bandwidth. With compatible hardware and software, this can make inference faster and cheaper. Preserving quality at such low precision remains a separate challenge.
As these techniques improve, increasingly capable models could fit on our own devices. This model is more or less GPT-4 level, so state of the art from two years ago, and really close to Opus 4.6, the frontier at the beginning of the year. At this pace, we’ll have Fable/Astra models by next year, more than enough to do most tasks most humans ask for.
Thus, I remain convinced that most AI tokens will eventually be generated locally, by smaller models handling most everyday tasks.
MODELS
Xiaomi clarifies that you can do a lot with little
Xiaomi has released MiMo-V2.6-Pro and the cheaper MiMo-V2.6-Flash, with Pro taking the top open-weight position on the Artificial Analysis Intelligence Index at 46 points. Pro supports text, images, audio, and video, with a one-million-token context window and downloadable weights under the MIT license.
Xiaomi says the models underwent a publicly documented reinforcement-learning run lasting less than six days, costing approximately $2.62 million for Pro and $850,000 for Flash, or $3.47 million combined. Each completed 30 training steps, learning from extended attempts at tasks such as coding and using software tools. These figures cover the disclosed reinforcement-learning run, rather than the models’ full development and pretraining costs.
The models were trained “in the open,” meaning Xiaomi shared the reinforcement-learning experiment live. The company says that the public run followed six months of research and engineering. Alongside the weights, it released a technical report, more than 7,000 training task environments, and an end-to-end reinforcement-learning framework, giving researchers resources to reproduce and extend the work.
Pro’s API costs $0.435 per million uncached input tokens and $0.87 per million output tokens. Xiaomi’s published evaluations show competitive results across coding and professional tasks, although Pro still trails Claude Opus 5 on tests including ProgramBench and Terminal Bench 4.0.
TheWhiteBox’s takeaway:
Lots of implications here. We can't compare total training costs between models from both countries, as Xiaomi does not publish the full training cost (and neither do American labs).
However, this run was probably in the low tens of millions for the golden training run (plus any additional training compute costs from experimentation; Xiaomi explicitly says those six days followed half a year of research and engineering experimentation, so those additional costs are definitely not negligible).
So this doesn't show that Chinese models are cheaper to train (although they obviously are because of differences in compute access), but it does show that a few million dollars of additional RL compute goes a long way.
This raises an interesting question: how much spending is necessary to remain competitive near the frontier? If one of the most critical parts just took under $3 million, one has to wonder whether frontier-level models really require $1 billion in training costs (the rumored total training cost for Astra).
If you ask me, it's becoming clear that the model architecture isn't a moat, and the real moat is your ability to get great training data.
That said, that doesn't automatically mean it's cheap. Obtaining data is a +$10B/year effort for US frontier labs because it requires significant compute to generate (i.e., synthetic data), plus high hourly rates for PhD-level human-generated data (e.g., via companies like Mercor, which have multiple billion ARR).
However, if data is so vital, this might shift the training cost formula away from compute spend and toward wages, where the US and China differ sharply. Thus, I wouldn't be surprised if one reason Chinese training costs are much lower is lower-cost access to data.
Seeing this, are we overbuilding compute? I'm not prepared to answer this question, as we need compute mostly for inference, so the spending could still be justified, and I'm still inclined to justify it. However, these results don't clearly show that we need to spend a single-digit % of GDP on compute.
On a final note, the fact that they don't mention the GPU provider makes it very likely they relied on US compute. For now, Chinese XPUs seem to work really well for inference (confirmed by Z.ai, DeepSeek...), but they are unproven for frontier-level training.
To end with a hot take: getting great results on low-single-digit millions could pave the way for corporations with strong IT budgets (basically every large corp) to seriously consider starting RL training on their own models using US open models. Law firm Latham & Watkins is purchasing GPUs, so it's not just technology and trading companies (e.g., Jane Street) doing this already.
Could all enterprises eventually become "labs"?
Personally, I believe the answer is 'yes'. In the future, 'adoption' might not mean using third-party model APIs and building your workflows on top, but actually taking strong, capable models and retraining them on your workflows—pretty much what Bridgewater is doing.
I largely agree with how Palantir/Microsoft CEOs describe this: you need to aspire to control your AI stack to remain competitive. So this could be a strong metric for investors to decide who will be the winners and losers in the adoption layer.
Closing Thoughts
As they say, there are decades when nothing happens, and weeks when decades pass. AI feels like that many times.
Even if one axis of progress stalls, many others continue to improve: whether that is model compression, training efficiency, or adoption-layer companies getting more sophisticated.
In many ways, the industry is in its infancy, so this isn’t surprising.
There’s still a lot that could go wrong (debt and weak revenues are always in the back of my mind, even if we haven’t), but a lot is going right right now.
But the obvious highlight is Jev, a model that can completely change the dynamics of AI adoption. Cheap, ultra-fast, and generalizable; a model that just works and is much more adoptable for companies because it doesn’t force companies to ‘re-do’ their business workflows, which is what LLMs require; it just enhances them.
Jev isn't mind-bending from a technological perspective; it’s just an ol’ reliable machine learning classifier. But it has already changed how we approach AI adoption. In a way, it has already changed the rules of the game.
Until next time!

Closing Thoughts

Give a Rating to Today's Newsletter
For business inquiries, reach me out at [email protected]
